@niw

Software engineering, loving cats.

github.com/niw
Joined April 2007
  • Tweets87.2K
  • Following2.5K
  • Followers72.2K
6.4KMedia
I recently made my 60% ANSI layout keyboard, mfk60 and here is the open source repository that contains all CAD projects and tooling source codes! github.com/niw/mfk60

GitHub - niw/mfk60: 60% ANSI Layout Split Keyboard. mfk60 is My First Keyboard, 60%.

60% ANSI Layout Split Keyboard. mfk60 is My First Keyboard, 60%. - niw/mfk60

github.com
2
2
33
21,483
macOS 27 でカラーピッカーのスポイトツールで自身のウィンドウから色を選ぶと WindowServer がフリーズして完全に何もできなくなるの面白すぎる。 WindowServer プロセスをなんとか再起動できれば戻せるんだけど難しいよね。
1
1,650
It is feeling pretty good when they refund for the overcharged Clipper card payment. But, it would be much better if the system work as expected to not overcharge at all, upfront. I know making system that works 80% not 99% would drastically reduces costs and time though...
2
1,865
実験的な --draft-steps ってのをつけた。4-step LoRA で最初の step を低解像度でつくってあとは高解像度ってしても動画生成できて30%くらい速くなるっていうやつ。DGX Spark で5秒 768p が70秒くらいだった設定で、50秒くらいになった。まあでもちょっと厳しい感じはある。 github.com/niw/mmh3
1
5
2,035
bilinearじゃなくて、ちゃんと upscaler を導入してだいぶマシになった。
1
2
518
canImport(SwiftUI, _version: 8.0.85.27)
1,853
みんなやってるから、一応、実装させてみた。Clef でマリオをプレイを MacBook で。
1
12
2,352
Added FL2VA and Ref2VA for Metal, tested with well-known images in macOS. It takes about a minute on M6 device. github.com/niw/mmh3
2
1,833
Result is kind of impressive, M6 with limited 24GB memory which has Neural Accelerators can generate video 2x faster than M4 Max by streaming model from SSD. Updated code and some benchmark results is here: github.com/niw/mmh3

GitHub - niw/mmh3: An inference engine dedicated to MiniMax H3.

An inference engine dedicated to MiniMax H3. Contribute to niw/mmh3 development by creating an account on GitHub.

github.com
Yoshimasa Niwa
@niwSep 26
After optimizing Metal Performance Primitives for Mac GPUs, if it can use Neural Accelerators like M5 and later, now it can generate short video about 3 minutes. Still working on it... github.com/niw/mmh3
1
10
3,730
Of course, CUDA is a lot faster, an order of magnitude faster, though it uses a lot of power. Probably performance per wattage would be a key at some point near future.
1
2
882
Now it works on RTX 4090 as well.
520
After optimizing Metal Performance Primitives for Mac GPUs, if it can use Neural Accelerators like M5 and later, now it can generate short video about 3 minutes. Still working on it... github.com/niw/mmh3
1
18
5,422
How can we prevent Xcode from downloading documentations always after stopping it?
1
2,120
May be this. defaults write com\.apple\.dt\.Xcode IDEDeveloperDocumentationDownloadableEnabled -bool NOdefaults write com\.apple\.dt\.Xcode IDEDeveloperDocumentationSelectedDuringFirstLaunch -bool NO
494
MiniMax H3 専用の推論実装、分散推論に対応した。例えば DGX Spark と MacBook Pro を両方を使って動画生成とかできるようになった。なお、Metal では遅すぎて意味はない模様... とはいえ、DGX Spark や RTX 5090 とかが複数ある環境なら速くなるはず。 github.com/niw/mmh3

GitHub - niw/mmh3: An inference engine dedicated to MiniMax H3.

An inference engine dedicated to MiniMax H3. Contribute to niw/mmh3 development by creating an account on GitHub.

github.com
1
1
4
2,050
Amdahl の法則(あっ、これ進研ゼミでやったところだ！)に支配されるので、普通に独立して並列で推論したほうが単独の推論を分散させるよりは全体では速い。なので、単独で長いの作る、みたいな時じゃないとあんまり意味はない。
1
619
Jevくんと将棋 (API課金きれたら止まると思う) shogi-one.niw.app
1
5
3
24
5,061
話題の TypeSafe AI の System One で将棋対戦させてみているけれど爆速なのはさておき強いかどうかは判断できないな...
1
1
3
1,954
Added Metal backend for Mac. It works but, as expected, slow. It takes about a minute for small short 2s 256p video generation on M4 device. M5 or later may be faster but I don’t have it so 🤷 github.com/niw/mmh3

GitHub - niw/mmh3: An inference engine dedicated to MiniMax H3.

An inference engine dedicated to MiniMax H3. Contribute to niw/mmh3 development by creating an account on GitHub.

github.com
1
1,720
ローカルの動画生成モデル MiniMax H3 専用での推論を Rust と CUDA で自前で実装中。sm120f ターゲットでまあまあ速いんじゃないのレベルまで来た。fl2va と ref2va にも対応させた。この ref2va は 8-step LoRA と NVFP4 使って GB10で120秒。
1
7
2,453
with NVFP4, and 480p, now it can generate 5s video in 32s on GB10. But, quality... it’s probably not reasonable to buy the speed as long as it can reach real-time.
Yoshimasa Niwa
@niwSep 15
Added major few-steps patchs and LoRA supports, with TaoMate-H3, it can generate 768p 5s video in about a minute on GB10. github.com/niw/mmh3

GitHub - niw/mmh3: An inference engine dedicated to MiniMax H3.

An inference engine dedicated to MiniMax H3. Contribute to niw/mmh3 development by creating an account on GitHub.

github.com
2,328
Load more